Enhancing Aerospace Classified Information Security through Large-scale Models
Regarding the stringent requirements of information confidentiality review in the aerospace field,current manual screening methods are suffering from high costs and insufficient accuracy of keyword matc-hing.An enhanced review framework integrated with large language models is proposed to improve the effi-ciency and accuracy of confidential information screening.Initially,the characteristics of confidential infor-mation are analyzed in the aerospace sector,an architecture that enhances the auditing performance of large language models is introduced in this study,which is combined with dynamic domain-specific expert system prompts to enhance the granularity and accuracy of reviews among multiple perspectives including technical and business confidentiality.By introducing a dynamic system prompt mechanism,the framework is effec-tively combined the semantic understanding capabilities of large language models with the real-time upda-ting of keywords.Additionally,in order to prevent excessive auditing by the large language model,a hy-brid cross-training strategy is developed,which significantly improves the recall rate of confidential informa-tion that reaches by 96％.Experiments on a self-developed high quality test set of1000 entries demonstrates that the proposed method outperforms global open-source large language models by 18％in aerospace classi-fied information inspection tasks.
Large language modelText content moderationIntelligent agentFine-tuningAerospace classified information inspection
NETL
NSTL
万方数据
